NARX: Contribution ‐ factor‐based short‐term multinodal load forecasting for smart grid

Abstract

A novel multinodal load forecasting method is presented in this paper, which uses smart metered data available from a real-life distribution grid at the NIT Patna campus. Because of the dynamic nature of the load, individual loads need to be predicted simultaneously so that they represent the load of the same instant. The proposed method uses single-stage multinodal forecasting with and without the load contribution factor (LCF), thus reducing the complexity compared to existing two-stage multinodal forecasting methods while improving the forecasting accuracy. It utilizes a nonlinear autoregressive neural network model with exogenous input (NARX-NN), which uses its own predicted output as an input during forecasting; this improves the accuracy of the model and makes it less dependent on external input data compared to other variations of NN. The experimental results show that the proposed method outperforms the existing approaches for multinodal load forecasting of the practical distribution system under consideration. Under different input dataset scenarios, the average mean absolute percentage error (MAPE) of the proposed model is 1.44, which represents the best forecasting performance among the competing models. K E Y W O R D Sload contribution factor, multinodal load forecasting, NARX-NN model, short-term load forecasting, smart metering List of Symbols and Abbreviations: L n (h), load of n th node at hour h; TL(h), global load of the system at hour h; x, actual load of a node at a given instant; x min , minimum value of load of a node; x max , maximum value of load of a node; x new , normalized, load of a node at a given instant; y, variable to be predicted; x, exogenous variables used to predict y; n, m, number of time delays in input side, output side; φ, linear activation function for target variable; ;, non-linear activation function for hidden neurons; α 0 , output bias; α i , weights of the output layer; γ i0 , input bias; γ ij , input weights; i, index of n neurons; j, index of m inputs; δ ir , weight of feedback h r, t − 1 which is delayed by time t-1; R, Regression coefficient; y i , actual load; y͠ i , forecasted load; N, number of samples used for testing; AMI, Advanced metering infrastructure; ANN, Artificial neural network; ARIMA, Autoregressive integrated moving average; AH, All hour load; ACF